Hard limits for marketing agents: what a Forbidden Rule is for
Hard limits for marketing agents are deny rules that stop a call before it is sent — whoever or whatever made it — and they complement approvals rather than replace them. A Forbidden Rule says what an agent may never read, change or delete, for whom, and on which data.
Approvals are not limits
An approval gate asks a person about a specific change. That works until the person is busy, the run is unattended, or the change arrives through a path nobody was watching. Some things should not be asked about at all: the client account you only report on, the audience lists legal has not cleared, the billing settings nobody on the team should change.
Those are limits, and a limit has to hold whether or not anyone is looking.
Allow-lists and deny rules do different jobs
- An allow-list — which platforms an API key exposes, which accounts a task selects — says what a caller is given. It is the right tool for scoping an integration.
- A deny rule says what must never happen, whoever asks. It survives a new key, a new schedule, a new teammate and an account that appeared in the list overnight.
- Use both. Scope each caller narrowly, and keep a small set of deny rules for the things that are never acceptable.
Three dimensions a rule needs
- Action — reading, changing or deleting. "Reports only on this account" blocks changes and deletes but leaves reads alone.
- Who — people by role (admin, editor, member) and kinds of caller: people in the app, schedules and loops, API keys, connected apps. An agency might let its own team read a client account but block every connected app.
- What data — audiences and customer lists, personal data, conversions, budgets and bids, creatives, reports, billing, account settings, email, catalog. Data categories are sturdier than tool names: a rule about audiences keeps working when a new audience tool ships.
Test a rule before you turn it on
- 1Describe the rule in words and let a draft map it onto your real platforms and accounts — then read what it actually selected. A tool name that does not exist should be dropped and said, not stored.
- 2Simulate a few calls against it: one it must block, one it must allow.
- 3Save it off, turn it on deliberately, and watch the blocked-call count over the next days. A rule that never blocks anything may be aimed at the wrong thing.
Why every lane must ask
A limit enforced only in the chat window is a suggestion. Agents reach platforms through scheduled runs, autonomous loops, alert checks, dashboards, MCP clients, REST integrations, raw API requests and writes approved hours after they were queued. If one of those skips the check, the rule fails exactly where nobody is watching.
The test of a hard limit is simple: a blocked call should send nothing, on every path, and say which rule stopped it.
Frequently asked questions
- What is the difference between an approval and a Forbidden Rule?
- An approval asks a person about one change. A Forbidden Rule blocks a kind of call outright — on chosen platforms, accounts, tools or data, for chosen callers — before anything is sent, whether or not anyone is watching.
- Should I block by tool name or by data category?
- Prefer data categories and actions. A tool-specific rule covers only that tool, while a rule about audiences or budgets keeps covering new tools and other paths that touch the same data.
- Do Forbidden Rules apply to API keys and MCP clients?
- In Agent Planners, yes: the same check runs on MCP, the REST API and raw API requests, and a rule can be written specifically for API keys or connected apps.